Active Defense Guidance for Spacecraft in Multi-Strategy Engagement with Incomplete Information

This paper investigates the active defense guidance problem for spacecraft in a multi-strategy target–pursuer–defender scenario, where a target spacecraft with an active defender attempts to evade the pursuer. Considering the practical applications, the pursuer may adopt variable possible interception strategies, and it brings the multi-strategy engagement problem for the target spacecraft. The active defense guidance not only contends with such engagement uncertainties but also needs to handle the incomplete information and observation noise during the whole process. By modeling the scenario as a multiple partially observable Markov decision process, this paper proposes a spacecraft active defense guidance based on an adaptive dueling double deep Q-network (AD 3 QN) reinforcement learning algorithm. To effectively address the partial observation problem caused by multi-strategy engagement, the AD 3 QN algorithm utilizes a fusion structure of convolutional and recurrent networks that enables the extraction of fragmented information characteristics and time-dependent data modeling. Moreover, a mechanism, which can stack the input observations and reshape the buffer arrangement, is designed to process the incomplete information. Based on this, the networks extract state features and trajectory characteristic tensors from the stacked time series observations, enhancing the agent’s estimation and identification abilities. With the proposed AD 3 QN-based guidance method, the target spacecraft exhibits robust performance even in the presence of incomplete information and multi-strategy engagement. Finally, the performance of the AD 3 QN-based policy is evaluated through numerical experiments.